jan/opencl_fdfd
2
1
Fork 1
Code Issues Pull Requests Releases Wiki Activity

refactor solver (untested)

Browse Source
This commit is contained in:
jan 2016-08-03 14:51:25 -07:00
parent 8682ee1ff8
commit ff3951ba35
5 changed files with 195 additions and 125 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
opencl_fdfd/kernels/common.cl Normal file
View File

@ -0,0 +1,52 @@
/* Common code for E, H updates
*
* Template parameters:
* ctype string denoting type for storing complex field values
* shape list of 3 ints specifying shape of fields
*/
//Defines to clean up operation names
#define ctype {{ctype}}_t
#define zero {{ctype}}_new(0.0, 0.0)
#define add {{ctype}}_add
#define sub {{ctype}}_sub
#define mul {{ctype}}_mul
// Field sizes
const int sx = {shape[0]};
const int sy = {shape[1]};
const int sz = {shape[2]};
//Since we use i to index into Ex[], E[], ... rather than E[], do nothing if
// i is outside the bounds of Ex[].
if (i >= sx * sy * sz) {
PYOPENCL_ELWISE_CONTINUE;
}
// Given a linear index i and shape (sx, sy, sz), defines x, y, and z
// as the 3D indices of the current element (i).
// (ie, converts linear index [i] to field indices (x, y, z)
const int z = i / (sx * sy);
const int y = (i - z * sx * sy) / sx;
const int x = (i - y * sx - z * sx * sy);
// Calculate linear index offsets corresponding to offsets in 3D
// (ie, if E[i] <-> E(x, y, z), then E[i + diy] <-> E(x, y + 1, z)
const int dix = 1;
const int diy = sx;
const int diz = sx * sy;
//Pointer offsets into the components of a linearized vector-field
// (eg. Hx = H + XX, where H and Hx are pointers)
const int XX = 0;
const int YY = sx * sy * sz;
const int ZZ = sx * sy * sz * 2;
//Define pointers to vector components of each field (eg. Hx = H + XX)
__global ctype *Ex = E + XX;
__global ctype *Ey = E + YY;
__global ctype *Ez = E + ZZ;
__global ctype *Hx = H + XX;
__global ctype *Hy = H + YY;
__global ctype *Hz = H + ZZ;

77
opencl_fdfd/kernels/e2h.cl
View File

@ -1,17 +1,39 @@
/*
*
* H update equations
*
* Template parameters:
* mu False if (mu == 1) everywhere
* pmc False if no PMC anywhere
* common_cl Rendered code from common.cl
*
* Arguments:
* ctype *E E-field
* ctype *H H-field
* ctype *inv_mu 1/mu (at H-field locations)
* char *pmc Boolean mask denoting presence of PMC (at H-field locations)
* ctype *inv_dex 1/dx_e (complex cell widths for x direction at E locations)
* ctype *inv_dey 1/dy_e (complex cell widths for y direction at E locations)
* ctype *inv_dez 1/dz_e (complex cell widths for z direction at E locations)
*
*/
//Define sx, x, dix (and y, z versions of those)
{{dixyz_source}}
{{common_cl}}
//Define vectorized fields and pointers (eg. Hx = H + XX)
{{vec_source}}
__global ctype *inv_mu_x = inv_mu + XX;
__global ctype *inv_mu_y = inv_mu + YY;
__global ctype *inv_mu_z = inv_mu + ZZ;
__global ctype *pmc_x = pmc + XX;
__global ctype *pmc_y = pmc + YY;
__global ctype *pmc_z = pmc + ZZ;
// Wrap indices if necessary
/*
* Implement periodic boundary conditions
*
* ipx gives the index of the adjacent cell in the plus-x direction ([i]ndex [p]lus [x]).
* In the event that we start at x == (sx - 1), we actually want to wrap around and grab the cell
* where x == 0 instead, ie. ipx = i - (sx - 1) * dix .
*/
int ipx, ipy, ipz;
if ( x == sx - 1 ) {
ipx = i - (sx - 1) * dix;
@ -32,53 +54,56 @@ if ( z == sz - 1 ) {
}
//Update H components; set them to 0 if PMC is enabled there.
// Also divide by mu only if requested.
//Update H components; set them to 0 if PMC is enabled at that location.
//Mu division and PMC conditional are only included if {{mu}} and {{pmc}} are true
{% if pmc -%}
if (pmc[XX + i] != 0) {
Hx[i] = cdouble_new(0.0, 0.0);
if (pmc_x[i] != 0) {
Hx[i] = zero;
} else
{%- endif -%}
{
cdouble_t Dzy = cdouble_mul(cdouble_sub(Ez[ipy], Ez[i]), inv_dey[y]);
cdouble_t Dyz = cdouble_mul(cdouble_sub(Ey[ipz], Ey[i]), inv_dez[z]);
ctype Dzy = mul(sub(Ez[ipy], Ez[i]), inv_dey[y]);
ctype Dyz = mul(sub(Ey[ipz], Ey[i]), inv_dez[z]);
ctype x_curl = sub(Dzy, Dyz);
{%- if mu -%}
Hx[i] = cdouble_mul(inv_mu[XX + i], cdouble_sub(Dzy, Dyz));
Hx[i] = mul(inv_mu_x[i], x_curl);
{%- else -%}
Hx[i] = cdouble_sub(Dzy, Dyz);
Hx[i] = x_curl;
{%- endif %}
}
{% if pmc -%}
if (pmc[YY + i] != 0) {
Hy[i] = cdouble_new(0.0, 0.0);
if (pmc_y[i] != 0) {
Hy[i] = zero;
} else
{%- endif -%}
{
cdouble_t Dxz = cdouble_mul(cdouble_sub(Ex[ipz], Ex[i]), inv_dez[z]);
cdouble_t Dzx = cdouble_mul(cdouble_sub(Ez[ipx], Ez[i]), inv_dex[x]);
ctype Dxz = mul(sub(Ex[ipz], Ex[i]), inv_dez[z]);
ctype Dzx = mul(sub(Ez[ipx], Ez[i]), inv_dex[x]);
ctype y_curl = sub(Dxz, Dzx);
{%- if mu -%}
Hy[i] = cdouble_mul(inv_mu[YY + i], cdouble_sub(Dxz, Dzx));
Hy[i] = mul(inv_mu_y[i], y_curl);
{%- else -%}
Hy[i] = cdouble_sub(Dxz, Dzx);
Hy[i] = y_curl;
{%- endif %}
}
{% if pmc -%}
if (pmc[ZZ + i] != 0) {
Hz[i] = cdouble_new(0.0, 0.0);
if (pmc_z[i] != 0) {
Hz[i] = zero;
} else
{%- endif -%}
{
cdouble_t Dyx = cdouble_mul(cdouble_sub(Ey[ipx], Ey[i]), inv_dex[x]);
cdouble_t Dxy = cdouble_mul(cdouble_sub(Ex[ipy], Ex[i]), inv_dey[y]);
ctype Dyx = mul(sub(Ey[ipx], Ey[i]), inv_dex[x]);
ctype Dxy = mul(sub(Ex[ipy], Ex[i]), inv_dey[y]);
ctype z_curl = sub(Dyx, Dxy);
{%- if mu -%}
Hz[i] = cdouble_mul(inv_mu[ZZ + i], cdouble_sub(Dyx, Dxy));
Hz[i] = mul(inv_mu_z[i], z_curl);
{%- else -%}
Hz[i] = cdouble_sub(Dyx, Dxy);
Hz[i] = z_curl;
{%- endif %}
}

80
opencl_fdfd/kernels/h2e.cl
View File

@ -1,17 +1,45 @@
/*
*
* E update equations
*
* Template parameters:
* pec False if no PEC anywhere
* common_cl Rendered code from common.cl
*
* Arguments:
* ctype *E E-field
* ctype *H H-field
* ctype *oeps omega*epsilon (at E-field locations)
* ctype *Pl Entries of (diagonal) left preconditioner matrix
* char *pec Boolean mask denoting presence of PEC (at E-field locations)
* ctype *inv_dhx 1/dx_h (complex cell widths for x direction at H locations)
* ctype *inv_dhy 1/dy_h (complex cell widths for y direction at H locations)
* ctype *inv_dhz 1/dz_h (complex cell widths for z direction at H locations)
*
*/
//Define sx, x, dix (and y, z versions of those)
{{dixyz_source}}
//Define vectorized fields and pointers (eg. Hx = H + XX)
{{vec_source}}
{{common_cl}}
// Wrap indices if necessary
__global ctype *oeps_x = oeps + XX;
__global ctype *oeps_y = oeps + YY;
__global ctype *oeps_z = oeps + ZZ;
__global ctype *pec_x = pec + XX;
__global ctype *pec_y = pec + YY;
__global ctype *pec_z = pec + ZZ;
__global ctype *Pl_x = Pl + XX;
__global ctype *Pl_y = Pl + YY;
__global ctype *Pl_z = Pl + ZZ;
/*
* Implement periodic boundary conditions
*
* imx gives the index of the adjacent cell in the minus-x direction ([i]ndex [m]inus [x]).
* In the event that we start at x == 0, we actually want to wrap around and grab the cell
* where x == (sx - 1) instead, ie. imx = i + (sx - 1) * dix .
*/
int imx, imy, imz;
if ( x == 0 ) {
imx = i + (sx - 1) * dix;
@ -34,38 +62,38 @@ if ( z == 0 ) {
//Update E components; set them to 0 if PEC is enabled there.
{% if pec -%}
if (pec[XX + i] == 0)
if (pec_x[i] == 0)
{%- endif -%}
{
cdouble_t tEx = cdouble_mul(Ex[i], oeps[XX + i]);
cdouble_t Dzy = cdouble_mul(cdouble_sub(Hz[i], Hz[imy]), inv_dhy[y]);
cdouble_t Dyz = cdouble_mul(cdouble_sub(Hy[i], Hy[imz]), inv_dhz[z]);
tEx = cdouble_add(tEx, cdouble_sub(Dzy, Dyz));
Ex[i] = cdouble_mul(tEx, Pl[XX + i]);
ctype tEx = mul(Ex[i], oeps_x[i]);
ctype Dzy = mul(sub(Hz[i], Hz[imy]), inv_dhy[y]);
ctype Dyz = mul(sub(Hy[i], Hy[imz]), inv_dhz[z]);
tEx = add(tEx, sub(Dzy, Dyz));
Ex[i] = mul(tEx, Pl_x[i]);
}
{% if pec -%}
if (pec[YY + i] == 0)
if (pec_y[i] == 0)
{%- endif -%}
{
cdouble_t tEy = cdouble_mul(Ey[i], oeps[YY + i]);
cdouble_t Dxz = cdouble_mul(cdouble_sub(Hx[i], Hx[imz]), inv_dhz[z]);
cdouble_t Dzx = cdouble_mul(cdouble_sub(Hz[i], Hz[imx]), inv_dhx[x]);
tEy = cdouble_add(tEy, cdouble_sub(Dxz, Dzx));
Ey[i] = cdouble_mul(tEy, Pl[YY + i]);
ctype tEy = mul(Ey[i], oeps_y[i]);
ctype Dxz = mul(sub(Hx[i], Hx[imz]), inv_dhz[z]);
ctype Dzx = mul(sub(Hz[i], Hz[imx]), inv_dhx[x]);
tEy = add(tEy, sub(Dxz, Dzx));
Ey[i] = mul(tEy, Pl_y[i]);
}
{% if pec -%}
if (pec[ZZ + i] == 0)
if (pec_z[i] == 0)
{%- endif -%}
{
cdouble_t tEz = cdouble_mul(Ez[i], oeps[ZZ + i]);
cdouble_t Dyx = cdouble_mul(cdouble_sub(Hy[i], Hy[imx]), inv_dhx[x]);
cdouble_t Dxy = cdouble_mul(cdouble_sub(Hx[i], Hx[imy]), inv_dhy[y]);
tEz = cdouble_add(tEz, cdouble_sub(Dyx, Dxy));
Ez[i] = cdouble_mul(tEz, Pl[ZZ + i]);
ctype tEz = mul(Ez[i], oeps_z[i]);
ctype Dyx = mul(sub(Hy[i], Hy[imx]), inv_dhx[x]);
ctype Dxy = mul(sub(Hx[i], Hx[imy]), inv_dhy[y]);
tEz = add(tEz, sub(Dyx, Dxy));
Ez[i] = mul(tEz, Pl_z[i]);
}
/*
* End H update equations
* End E update equations
*/

24
opencl_fdfd/kernels/p2e.cl
View File

@ -1,9 +1,29 @@
/*
* Apply PEC and preconditioner.
*
* Template parameters:
* ctype name of complex type (eg. cdouble)
* pec false iff no PEC anyhwere
*
* Arguments:
* ctype *E (output) E-field
* ctype *Pr Entries of (diagonal) right preconditioner matrix
* ctype *p (input vector)
*
*/
//Defines to clean up operation names
#define ctype {{ctype}}_t
#define zero {{ctype}}_new(0.0, 0.0)
#define mul {{ctype}}_mul
{%- if pec -%}
if (pec[i] != 0) {
E[i] = cdouble_new(0.0, 0.0);
E[i] = zero;
} else
{%- endif -%}
{
E[i] = cdouble_mul(Pr[i], p[i]);
E[i] = mul(Pr[i], p[i]);
}

87
opencl_fdfd/ops.py
View File

@ -29,65 +29,10 @@ def type_to_C(float_type: numpy.float32 or numpy.float64) -> str:
return types[float_type]
def shape_source(shape) -> str:
"""
Defines sx, sy, sz C constants specifying the shape of the grid in each of the 3 dimensions.
:param shape: [sx, sy, sz] values.
:return: String containing C source.
"""
sxyz = """
// Field sizes
const int sx = {shape[0]};
const int sy = {shape[1]};
const int sz = {shape[2]};
""".format(shape=shape)
return sxyz
# Defines dix, diy, diz constants used for stepping in the x, y, z directions in a linear array
# (ie, given Ex[i] referring to position (x, y, z), Ex[i+diy] will refer to position (x, y+1, z))
dixyz_source = """
// Convert offset in field xyz to linear index offset
const int dix = 1;
const int diy = sx;
const int diz = sx * sy;
"""
# Given a linear index i and shape sx, sy, sz, defines x, y, and z
# as the 3D indices of the current element (i).
xyz_source = """
// Convert linear index to field index (xyz)
const int z = i / (sx * sy);
const int y = (i - z * sx * sy) / sx;
const int x = (i - y * sx - z * sx * sy);
"""
vec_source = """
if (i >= sx * sy * sz) {
PYOPENCL_ELWISE_CONTINUE;
}
//Pointers into the components of a vectorized vector-field
const int XX = 0;
const int YY = sx * sy * sz;
const int ZZ = sx * sy * sz * 2;
"""
E_ptrs = """
__global cdouble_t *Ex = E + XX;
__global cdouble_t *Ey = E + YY;
__global cdouble_t *Ez = E + ZZ;
"""
H_ptrs = """
__global cdouble_t *Hx = H + XX;
__global cdouble_t *Hy = H + YY;
__global cdouble_t *Hz = H + ZZ;
"""
preamble = '''
#define PYOPENCL_DEFINE_CDOUBLE
#include <pyopencl-complex.h>
'''
ctype = type_to_C(numpy.complex128)
@ -98,15 +43,17 @@ def ptrs(*args):
def create_a(context, shape, mu=False, pec=False, pmc=False):
header = shape_source(shape) + dixyz_source + xyz_source
vec_h = vec_source + E_ptrs + H_ptrs
common_source = jinja_env.get_template('common.cl').render(shape=shape,
ctype=ctype)
pec_arg = ['char *pec']
pmc_arg = ['char *pmc']
des = [ctype + ' *inv_de' + a for a in 'xyz']
dhs = [ctype + ' *inv_dh' + a for a in 'xyz']
p2e_source = jinja_env.get_template('p2e.cl').render(pec=pec)
p2e_source = jinja_env.get_template('p2e.cl').render(pec=pec,
ctype=ctype)
P2E_kernel = ElementwiseKernel(context,
name='P2E',
preamble=preamble,
@ -115,8 +62,7 @@ def create_a(context, shape, mu=False, pec=False, pmc=False):
e2h_source = jinja_env.get_template('e2h.cl').render(mu=mu,
pmc=pmc,
dixyz_source=header,
vec_source=vec_h)
common_cl=common_source)
E2H_kernel = ElementwiseKernel(context,
name='E2H',
preamble=preamble,
@ -124,8 +70,7 @@ def create_a(context, shape, mu=False, pec=False, pmc=False):
arguments=', '.join(ptrs('E', 'H', 'inv_mu') + pmc_arg + des))
h2e_source = jinja_env.get_template('h2e.cl').render(pec=pec,
dixyz_source=header,
vec_source=vec_h)
common_cl=common_source)
H2E_kernel = ElementwiseKernel(context,
name='H2E',
preamble=preamble,
@ -143,8 +88,8 @@ def create_a(context, shape, mu=False, pec=False, pmc=False):
def create_xr_step(context):
update_xr_source = '''
x[i] = cdouble_add(x[i], cdouble_mul(alpha, p[i]));
r[i] = cdouble_sub(r[i], cdouble_mul(alpha, v[i]));
x[i] = add(x[i], mul(alpha, p[i]));
r[i] = sub(r[i], mul(alpha, v[i]));
'''
xr_args = ', '.join(ptrs('x', 'p', 'r', 'v') + [ctype + ' alpha'])
@ -191,7 +136,7 @@ def create_rhoerr_step(context):
def create_p_step(context):
update_p_source = '''
p[i] = cdouble_add(r[i], cdouble_mul(beta, p[i]));
p[i] = add(r[i], mul(beta, p[i]));
'''
p_args = ptrs('p', 'r') + [ctype + ' beta']
@ -214,9 +159,9 @@ def create_dot(context):
name='dot',
preamble=preamble,
dtype_out=dot_dtype,
neutral='cdouble_new(0.0, 0.0)',
map_expr='cdouble_mul(p[i], v[i])',
reduce_expr='cdouble_add(a, b)',
neutral='zero',
map_expr='mul(p[i], v[i])',
reduce_expr='add(a, b)',
arguments=ptrs('p', 'v'))
def ri_update(p, v, e):
@ -230,14 +175,14 @@ def create_a_csr(context):
spmv_source = '''
int start = m_row_ptr[i];
int stop = m_row_ptr[i+1];
cdouble_t dot = cdouble_new(0.0, 0.0);
dtype dot = zero;
int col_ind, d_ind;
for (int j=start; j<stop; j++) {
col_ind = m_col_ind[j];
d_ind = j;
dot = cdouble_add(dot, cdouble_mul(v_in[col_ind], m_data[d_ind]));
dot = add(dot, mul(v_in[col_ind], m_data[d_ind]));
}
v_out[i] = dot;
'''